MEML-GRPO: Heterogeneous Multi-Expert Mutual Learning for RLVR Advancement

TL;DR

This paper introduces MEML-GRPO, a novel multi-expert mutual learning framework for reinforcement learning with verifiable rewards, significantly improving reasoning performance of large language models by addressing reward sparsity and enhancing knowledge sharing.

Contribution

The paper presents a new multi-expert mutual learning approach that uses diverse prompts and inter-expert knowledge transfer to improve RLVR in large language models.

Findings

Achieves an average performance gain of 4.89% with Qwen.

Achieves an average performance gain of 11.33% with Llama.

Effectively overcomes reward sparsity issues in RLVR.

Abstract

Recent advances demonstrate that reinforcement learning with verifiable rewards (RLVR) significantly enhances the reasoning capabilities of large language models (LLMs). However, standard RLVR faces challenges with reward sparsity, where zero rewards from consistently incorrect candidate answers provide no learning signal, particularly in challenging tasks. To address this, we propose Multi-Expert Mutual Learning GRPO (MEML-GRPO), an innovative framework that utilizes diverse expert prompts as system prompts to generate a broader range of responses, substantially increasing the likelihood of identifying correct solutions. Additionally, we introduce an inter-expert mutual learning mechanism that facilitates knowledge sharing and transfer among experts, further boosting the model's performance through RLVR. Extensive experiments across multiple reasoning benchmarks show that MEML-GRPO delivers significant improvements, achieving an average performance gain of 4.89% with Qwen and 11.33% with Llama, effectively overcoming the core limitations of traditional RLVR methods.